Search

[2610.37826] DeFT-RAG: Decomposing Multi-Hop Questions into Self-Contained Sub-Questions with Reusable Answers

DeFT-RAG improves multi-hop question answering by generating self-contained sub-questions that explicitly name involved entities, offloading remaining work to rule-based updates and a lightweight decision model, and reusing cached QA pairs to reduce LLM usage without sacrificing accuracy.

How to Teach an AI Assistant to Answer Questions Using a Company Knowledge Base / Habr

Translated title. Original: Как научить ИИ-ассистента отвечать на вопросы, опираясь на базу знаний компании? / Хабр

The text combines an announcement for a VK AI Space webinar on building a RAG assistant with an article excerpt reporting security and privacy issues found in an audit of 284 Russian company websites.

Digest by Avi Flombaum

The piece describes building an eval and tracing workflow to test AI apps, using real traffic questions and repeated runs to detect when fixes break other behavior; it emphasizes turning real failures into repeatable tests to understand changes across releases.

DeepSeek Opens App Features Through Its API — CASRAI

CASRAI promotes a research-administration AI assistant that cites sources and draws on its guides, dictionary, and indexed government and funder documents. The page lists free and paid question limits and says CASRAI publications remain free to read.

Qbim launches the AI-powered sales system Q-Sales 2.0

Translated title. Original: Qbim lanserar AI-baserade säljsystemet Q-Sales 2.0

Qbim launched Q-Sales 2.0, a new version of its AI-powered sales system, along with a new website. The article gives no further details about the product.

Bergen company builds AI solution for Brann – now they want to challenge the big ones

Translated title. Original: Bergensselskap bygger KI-løsning for Brann – nå vil de utfordre de store

A Bergen company builds an AI tool to help Brann with questions and ticketing; developers admit accuracy issues, and the full story is behind a subscription.

[2610.23400] On the Tip of the Tongue: First-Token Cueing as a Behavioral Probe of Recall Failures in LLMs

The paper shows that many closed-book errors in LLMs are first-token recall failures; a simple First-Token Cueing probe reveals that providing the first token of the correct answer reduces errors across multiple models and languages, indicating encoded knowledge and suggesting focus on the initial decoding step for improvement.

First harmonized standard to support the AI Regulation published in the Official Journal

Translated title. Original: Erste harmonisierte Norm zur Unterstützung der KI-Verordnung im Amtsblatt veröffentlicht

EU发布了实现AI Regulation的执行决定，EN 18286:2026确立了高风险AI系统的质量管理体系(QMS)并提供合规推定，是基于AI Regulation第17条的唯一AI-QMS标准；该标准在高风险AI系统要求正式适用前被视为重要里程碑，预计未来还将有更多协调标准。关于此的更多信息请参阅欧盟委员会的新闻稿（2026年10月19日）。

I am excited": What telephone AI can do | APOTHEKE ADHOC

Translated title. Original: „Ich bin begeistert“: Was Telefon-KI leisten kann | APOTHEKE ADHOC

A pharmacy uses an AI telephone assistant to handle routine customer calls, improving backoffice focus. The AI, exemplified by 'Steffi' from Flying Pharmacist, demonstrates professional, empathetic responses, with handoff to staff for complex issues and configurable activation timing; onboarding and integration with inventory and online orders are essential.

[2610.20809] A Self-Improving Agent that Grows its Own DSL: Oracle-Free Operator Induction for Exact Question Answering

An agent expands exact-question answering by self-generating operators within a fixed DSL, using agreement and a conservation rule instead of ground truth; the process increases answerable questions across multiple domains and remains sound across model families.